For decades, game studios have constructed some of the richest interactive environments humans have ever produced. Worlds with controllable physics, observable state, scripted and emergent behavior, action under consequence. They built them for players. They built them well. And as it turns out, they built exactly the kind of data the next generation of AI needs to learn how reality works.
Today, Lightspeed Venture Partners is leading Origin Lab's $8M seed round. Eniac Ventures, FPV Ventures, Seven Stars, and SV Angel are joining, along with angels from robotics, AI, gaming, and Google, including Twitch co-founder Kevin Lin and Cruise founder Kyle Vogt. TechCrunch's coverage is here. GamesBeat's is here.
If you've followed our writing this spring, this announcement is the next step in a thesis we've been making publicly for months. We wrote about Why We Started Origin Lab. We wrote about Artificial World Intelligence™ as a category. We wrote about The Scraping Era Is Over. We wrote A Memo to AI CEOs: Bad Data Is Not Cheap Data. The capital announced today is what turns that thesis into infrastructure.
The money goes to one thing: building the channel between game studios and the AI labs training the next generation of world models, video generation systems, and embodied agents.
Where AI is, and what it needs
Frontier AI is moving past language and static images. The next class of models needs data that reflects how the world actually works. Motion, physics, spatial structure, action, environment state, cause and effect. World models, video generators, robotics foundations, planning systems for embodied agents. All of them are constrained by the same thing: the quality, structure, and provenance of the data they train on.
Scraped video gives you pixels. It does not give you depth. It does not give you the keyboard input that produced the action on screen. It does not give you the physics state behind the collision. It does not give you a license you can train on without getting sued.
The serious labs are past the scraping question. They're asking the harder one: where does the data we actually need come from, and how do we get it legally, at scale, with structure?
The answer was hiding in game worlds
A game engine, by design, is exactly the kind of data world models need. Every frame has depth. Every action has an input. Every collision has a physics state. Every scene has structured metadata. Every player session is a record of an agent making decisions in a consequence-rich environment. None of that exists in scraped video. All of it exists in a running engine.
The problem was never supply. The problem was that no one had built the technical and commercial path to turn those engines into datasets AI labs could buy, train on, and audit.
What we built
Origin Lab is that path.
We work directly with video game publishers to license content at the source. We capture it through proprietary pipelines at the engine level, not the screen: video, audio, human inputs, camera trajectory, physics state, and scene annotations, all pulled directly from the engine and time-aligned to the same frame clock. We enrich every dataset with more than 20 metadata categories. We deliver it through one API, structured for query.
Today we have exclusive partnerships with more than twenty game publishers representing more than fifty titles, with more added every month. We are already under contract with a leading frontier AI lab.
We call this emerging category Artificial World Intelligence™. Licensed, structured data and systems for AI models that must understand, simulate, and interact with complex environments.
Research
Origin Lab isn't only a data delivery company. There's applied research underneath: generative world-building, synthetic player models, scene intelligence, world value scoring, and automated data quality. We're in active collaboration with researchers across academia and inside corporate AI labs, including teams at Oxford, with several joint papers in progress. The platform and the research reinforce each other.
How we built it
This market only works if it works for everyone in it. Game studios spent decades building these worlds and deserve to participate in the value they create for AI. AI researchers deserve data they can trust: licensed at the source, audit-ready, structured for training.
So consent, attribution, revenue share, and usage tracking are built into the pipeline from day one. Not as a legal layer bolted on top. As the operating system the platform runs on.
What's next
The seed accelerates four things. Deeper capture and enrichment technology. More publisher partnerships. Growth of the engineering and research teams. More research collaborations.
We're hiring across all of it.
Thank you
To Faraz Fatemi, who saw this coming a year ago, and Lightspeed for leading. To Eniac, FPV, Seven Stars, and SV Angel for joining. To the operators and angels backing us across robotics, AI, gaming, and Google: Kevin Lin, Kyle Vogt, Peter Sellis, David Benhaim, Artia Moghbel, John Howell, Michael Carney, and Greg Arrese. Each of you has built the kind of platform, product, or system that gave us the conviction this category needed a category-defining company.
To our first AI lab customer, our publisher partners, and the researchers who told us early what they actually needed: thank you. You shaped what this is.
If you're reading this
If you're an AI lab training world models, multimodal systems, or embodied agents, we'd like to talk.
If you're a publisher with a catalog ready to participate, we'd like to talk.
If you want to build this with us, we'd really like to talk.
Colin and Antoine